
Bug localization for Register-Transfer Level (RTL) designs is essential for improving hardware reliability and reducing debugging costs. Spectrum-based bug localization (SBBL) techniques are widely adopted because of their efficiency; however, their accuracy is often limited because they rely solely on statistical correlations between execution traces and test outcomes. Mutation-based bug localization (MBBL) improves localization accuracy by analyzing behavioral changes induced by mutations. Nevertheless, existing MBBL approaches depend on predefined syntactic mutation operators, which frequently generate unrealistic or low-quality mutations, thereby introducing noise and limiting their effectiveness in RTL debugging. To address these limitations, we propose SeMuLoc, a novel large language model (LLM)-guided mutation-based bug localization method for RTL designs. First, we apply program slicing to constrain mutation analysis to failure-relevant regions. Next, we introduce an LLM-guided semantic mutation generation strategy that produces realistic bug patterns consistent with RTL design characteristics. To ensure mutation validity, we further incorporate a semantic consistency constraint to filter invalid mutation candidates. In addition, we develop a mutation weighting scheme that integrates semantic counterfactual plausibility and behavioral impact of mutations to refine suspiciousness computation by differentiating the relative importance of different mutations. To evaluate the effectiveness of SeMuLoc, we conduct extensive experiments on 228 bug versions across 14 hardware projects and compare it with four state-of-the-art bug localization techniques. Experimental results demonstrate that SeMuLoc achieves better performance than all baselines. In particular, SeMuLoc improves the performance of the state-of-the-art SBBL and MBBL methods by an average of 94.84% and 84.86%, respectively, under the RImp metric.
Edge devices employed for Internet of Thing (IoT) are vulnerable to Power Analysis Attack (PAA) owing to their ease of access. Although the information is encrypted using the cryptographic cipher, Advanced Encryption Standard (AES), power side-channel monitoring enables recovery of the secret key employed for encryption. This work discusses the methodology to perform PAA on pre-silicon Application Specific Integrated Circuit (ASIC) and Field Programmable Gate Array (FPGA) platforms, and test the design’s resilience using standard security metrics. A novel countermeasure technique is presented which employs the famous mathematical problem, Collatz conjecture, used to generate random noise to inflict the AES power signature, aided by temporal dynamization of round operations to prevent aligning of traces by adversaries. The countermeasure incurs overheads of 18.63% and 11.28% for area and power, respectively, while preserving operational performance appropriate for low-resource IoT deployments. Furthermore, the proposed design remained resilient beyond 1 million CPA traces indicating a high degree of side-channel robustness, while simultaneously exhibiting strong resistance against template-based and Deep Learning (DL)-assisted profiled attacks.
Accurate and swift RTL timing estimation is crucial for early-stage timing optimization and design space exploration. Existing learning-driven methods fail to address the complexities introduced by the trending hierarchical design paradigm, specifically non-uniform input arrival times (NUIAT). In this paper, we present NUA-Timer++, a framework for predicting post-synthesis timing under NUIAT in hierarchical RTL blocks implemented on field-programmable gate arrays (FPGAs). NUA-Timer++ features a multi-scale timing dependency encoder (MTDE) with cross-stage criticality alignment loss (CCAL) to learn a timing correlation matrix explicitly aligned with actual post-synthesis critical path distributions. Building upon this, NUA-Timer++ generates comprehensive endpoint representations by: (i) encoding timing propagation behaviors, (ii) capturing endpoint-wise fanin-cone structures via an attention mechanism, and (iii) performing parallel extraction and explicit encoding of probable post-synthesis critical paths using a correlation-based greedy tree search and a Correlation-Aware Pathformer. Comprehensive experiments on 23 held-out open-source designs spanning 1.8K to 1.3 million nodes demonstrate the accuracy and scalability of our method compared with state-of-the-art RTL timing evaluators.
Accurate and scalable modeling of coaxial through-glass-via (c-TGV) packages is essential for signal-integrity analysis in complex three-dimensional systems. However, the computational cost of conventional SPICE simulation increases with line count and coupling complexity, while existing compact electronic device modeling approaches require manual formulation and learned surrogates are generally evaluated using small, configuration-specific datasets. This work introduces a SPICE-derived benchmark dataset spanning 5-line, 10-line, and 20-line c-TGV structures, comprising 500, 1000, and 1600 transient scenarios, respectively, with diverse aggressor-victim interactions. Two hybrid neural architectures are proposed: a parallel attention channeled transformer (PACT), which fuses gated multi-path temporal representations, and a local-global extractor (LGE), which combines multi-scale convolutional processing with global recurrent context. Twelve neural architectures are systematically evaluated across depths of 1–6 using configuration-specific accuracy, depth and data scaling, metric-weight sensitivity, cross-configuration transfer, and unseen output-cardinality tests. The results reveal a configuration-dependent transition in the preferred architecture. For the 5-line configuration, a one-layer LSTM provides the best overall waveform reconstruction, achieving an MAE of 0.012 ± 0.003 and R2 = 0.996 ± 0.001, whereas PACT gives the lowest t50 error of 13.2 ± 18.6 ps. This shallow-LSTM advantage does not scale to larger arrays. At 10 and 20 lines, LGE achieves the best performance across all seven reported waveform and electronic metrics, attaining MAE/MSE/R2 values of 0.032±0.001/0.002±0.000/0.979±0.001 and 0.055±0.003/0.007±0.001/0.949±0.004, with corresponding t50 errors of 12.1±1.6 ps and 36.7±24.3 ps, respectively. Across all six directed cross-configuration transfers, LGE achieves the lowest MAE and MSE and the highest R2 and Pearson correlation, maintaining R2 values of 0.7867–0.9375 without retraining and reducing MAE by 5.2%–68.7% relative to the strongest competing architecture. Furthermore, LGE remains ranked first for every one of the 230,230 evaluated metric-weight vectors at both 10 and 20 lines and preserves the dominant transient behavior for unseen 3-, 7-, and 15-line outputs. These results establish a scalable benchmark for c-TGV transient modeling and identify local-global feature extraction as a practical foundation for fast, configuration-adaptive signal-integrity prediction in next-generation heterogeneous integration.
The rapid deployment of face-to-face (F2F) hybrid bonding technology has fundamentally changed 3D integrated circuit design, enabling high-density interconnects and breaking the traditional constraint that requires functional blocks to be placed on a single die. However, existing design flows are limited to inter-module partitioning, which severely constrains the design space and prevents optimization of intra-module cell distribution across dies. To address this challenge, we propose a floorplan-to-placement flow, SoftClay3D, that performs intramodule partitioning and floorplanning with a true-3D floorplanning engine, followed by placement of standard cells within each module using a true-3D placement engine, thereby enabling optimization of both inter-module and intra-module wirelength. Experimental results on large-scale designs ranging from 0.85M to 24.3M instances demonstrate that our floorplan-to-placement flow achieves an average 32.2% wirelength reduction, 6.0% improvement in WNS, and 12.7% improvement in TNS compared to Hier-3D, and 29.1% wirelength reduction, 5.8% improvement in WNS, and 12.0% improvement in TNS compared to Corblivar, a state-of-the-art open-source multi-objective 3D floorplanner. These results confirm the effectiveness of the proposed flow as a complete and competitive design methodology for modern 3D IC design.